R Systems formed in 2003 as a a spin-off of the National Center for Supercomputing Applications (NCSA). One of a handful of small, on-demand HPC cloud providers founded in the early 2000s, R Systems has differentiated itself from the larger cloud public cloud vendors by providing tailored access to bare metal HPC clusters — and now, R Systems sees another way to differentiate itself.
HPCwire caught up with R Systems recently to learn about its new offering: an approach combining dedicated infrastructure with cloud bursting capabilities in one datacenter, contrasting with the more well-known cloud bursting model offered by major public cloud providers.
Overdeployment and underutilization
R Systems’ first customer for this new offering is Inphi, a company that specializes in the development of high-speed data movement interconnects using electronic design automation (EDA) computing. For years, Inphi had struggled with a problem common to EDA software users: Inphi’s computing needs peaked a few times a year, and it was buying hardware to meet those peak needs.
This meant that for much of the year, Inphi’s internally-owned computing hardware sat largely unused. The company said that previously, they were forced to purchase twice as much computing capacity as they generally needed, resulting in low utilization and high costs. “We at Inphi, like the rest of the industry, were well aware of this over-spending model,” said Scott Clark, Inphi’s vice president of IT, “but our options were limited.”
This led Inphi to seek alternatives to on-premises deployment. Cloud bursting – sending some peak workloads to a cloud service provider – had been a popular solution in similar use cases, as it eliminated the need to buy more hardware than was necessary for constant load. However, this hybrid on-premises/cloud approach posed its own drawbacks — in particular, high latency when transferring large amounts of data, resulting in slowdowns during critical crunch periods. Inphi also worried about the impacts on complexity and security that might come with a shift to operating both an on-premises system and a cloud system.
The ‘under one roof’ approach
R Systems laid the groundwork for a new hybrid cloud bursting model when it expanded its infrastructure into Switch, a Tier 5 platinum datacenter in Las Vegas. Inphi was already colocating its computing resources at facilities like these, and R Systems introduced the ability to do cloud bursting in the same facility where client-owned equipment was colocated.
For its initial “under one roof” deployment, Inphi’s available computing power comprised 10,000 owned cores and an additional 10,000 burst cores available for peak usage, connected by dark fiber.
Subsequently, Inphi’s dedicated cluster now sees utilization rates of 95 percent versus a previous low of 50 percent. The fact that the data never leaves the Switch facility or gets processed on a shared resource allays security concerns that Inphi had previously when considering hybrid cloud solutions.
When Inphi isn’t using those additional cores for cloud bursting, R Systems schedules other users into the gaps. Ideally, R Systems explained, its major users schedule resources around a year in advance, allowing them to optimize the load across its machines throughout the year – but it can also leverage resources in as little as one week. When resources remain available between scheduled usages, the company might offer availability at a discounted rate.
For now, R Systems’ computing power is driven primarily by dual-socket Intel configurations – including Skylake, Ivy Bridge and Broadwell CPUs – as well as some AMD- and GPU-powered systems. Michael Senizaiz, CTO of R Systems, explained that the largely homogeneous composition of their systems allows for greater scalability, which tended to benefit users more than the performance gains from running specialized hardware.
At the moment, the cost benefits of R Systems’ hybrid model are unclear in terms of hard numbers, though representatives suggested that it was cost-competitive, largely due to the savings gained from not having to overbuild hardware for peak bursting periods. The company pointed out other benefits of its model – such as increased security and decreased risk of producing a faulty product – are more difficult to quantify.
Room for growth
R Systems says it is currently in talks with three more EDA clients interested in the “under one roof” approach. Brian Kucic, principal and co-founder of R Systems, hopes that users who get a taste of the “under one roof” will be likely to buy in for greater usage over time. “It seems like with many of our users, the more you give them, the more they use,” he said. And, while R Systems is targeting the EDA space first, they are looking at expanding the service to other suitable industries — notably, oil and gas and manufacturing.
Inphi is similarly hopeful about R Systems’ approach. “If other EDA companies like Broadcom and Qualcomm adopt this model,” Clark said, “it’ll ensure more EDA-optimized infrastructure is available for bursting.”
“There is a lot of elegance in the simplicity of running all of our HPC workloads under one roof,” he continued. He likened it to the early days of 100 percent on-premises computing. “Simplicity,” he said, “is always better.”